Distributed model predictive control using sensitivity based primal decomposition
Third party funded individual grant
Start date : 01.01.2026
End date : 31.12.2028
Project details
Short description
Model Predictive Control (MPC) is a widely used strategy for controlling linear and nonlinear systems. It is based on the iterative solution of a dynamic optimization problem over a receding horizon. For networked and coupled systems, distributed MPC (DMPC) is an attractive extension of MPC, where the central MPC controller is replaced by local MPC agents for the individual subsystems of the global system. The probably most popular DMPC method is ADMM (Alternating Direction Method of Multipliers) that is based on the dual decomposition of the distributed problem.
A promising alternative approach, which has not yet been fully explored in the literature, is sensitivity-based primal decomposition. In this approach, the individual agents explicitly consider the costs of their actions on the neighbors’ performance. These sensitivities can be locally computed in an efficient mannery. Compared to ADMM, sensitivity-based DMPC shows an improved convergence behavior, reduced communication overhead, and lower algorithmic complexity. The efficient local computation of sensitivities and a simpler convergence analysis are further advantages of this method. Despite these advantages, the sensitivity-based approach currently has several shortcomings compared to ADMM. In particular, convergence and stability can only be guaranteed for a maximum prediction horizon and general state couplings are more difficult to consider with primal decomposition.
Therefore, this project aims to conduct an in-depth investigation of the sensitivity-based approach for distributed model predictive control. In particular, the aforementioned shortcomings compared to ADMM shall be addressed, and the overarching topic of sensitivities can be used to increase efficiency and flexibility in numerical solutions, to simplify the methodological analysis, and to enable practical implementation. The findings will be published in the DMPC toolbox GRAMPC-D.
